Intelligent agents that effectively adapt to changes in their task through novel architectures, training schemes, and evaluation methods.
Novel environments for exercising lifelong learning agents
Training and evaluating a lifelong learning agent requires the ability to generate a diverse variety of tasks and generate rich systematic variation with each task.
MetaArcade is a highly parameterized environment to construct custom 2D arcade games with controlled variation of perceptual features, actions and rewards.
References: https://arxiv.org/abs/2112.00583 and https://github.com/darpa-l2m/meta-arcade
L2Explorer is a Unity-based first-person-view 3D exploration environment that can be configured on the fly to generate a range of tasks and task variants which can be structured into complex and evolving evaluation curricula.
Novel algorithms to handle perceptual and task variation
Artificial agents struggle when asked to identify objects under different visual conditions, act upon input observations that are superficially different (e.g., different lighting conditions in natural imagery), or adapt to significant changes in their task.
Figure 1: The appearance of a ship and its surrounding context (waves, wakes) can vary significantly depending on weather, lighting, and orientation relative to the camera.